Urban Building Density Estimation from High-Resolution Imagery Using Multiple Features and Support Vector Regression.

Tao Zhang,Xin Huang,Dawei Wen,Jiayi Li

DOI: https://doi.org/10.1109/jstars.2017.2669217

IF: 4.715

2017-01-01

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing

Abstract:Building density is considered as an environmental parameter of great significance for urban management, and it can also help us to understand how cities function. However, in developing countries, building density information is often lacking or incomplete in a large number of cities. To address this problem, this paper presents a method for the quantitative estimation of building density using a support vector regression model to establish the relationship between building density and the features extracted from the image, including spectral, morphological, and textural features. The importance and relevance of the different features is investigated by a feature selection method named recursive feature elimination. Tests performed on three representative megacities in China confirm that the proposed method achieves satisfactory results for building density estimation in terms of the low root-mean-square errors (similar to 0.05) and visual effect. It is also found that the combination of feature sets outperforms the single feature set, indicating the need for simultaneous consideration of all the categories of features. The knowledge of the most informative features provides an insight into selecting the most effective parameters for building density mapping. The investigation of the transferability between geographical regions confirms that the self-training achieves the better performance compared to the transfer training. The results with regard to the regression accuracy, feature selection, and the comparison with other methods indicate the effectiveness of the multifeature approach when applied to building density estimation.